Wound severity assessment method and system based on artificial intelligence and big data analysis

Abstract

The invention provides a trauma severity assessment method and system based on artificial intelligence and big data analysis, and relates to the technical field of AI.The trauma severity assessment method comprises the steps that an AI model is guided through cue word engineering to extract trauma information from structured medical record data of a trauma patient, and JSON format data containing diagnosis information, trauma identification and trauma description is generated; and a vector retrieval technology is adopted to retrieve a scoring rule related to a single injury condition from the modified ISS scoring rule knowledge base, and the AI model determines the area, the part and the score of each injury condition based on the JSON data and the scoring rule. And after checking and combining the injury condition scores, finally outputting an ISS total score and a specific score result of each wound area. The method can automatically extract and score the injury condition information, improves the ISS scoring efficiency and accuracy, and assists a doctor in rapidly evaluating the injury condition of a patient.

Description

Technical Field

[0001] The present invention relates to AI technology, and in particular to a trauma severity assessment method and system based on artificial intelligence and big data analysis. Background Art

[0002] The assessment of injuries caused by trauma is crucial for the treatment of patients. Rapid and accurate assessment can guide medical intervention and improve prognosis. The Injury Severity Score (ISS) is a commonly used anatomical scoring system used to assess the overall injury severity of patients with multiple injuries. Traditional ISS scoring relies on manual review of scoring rules and calculations based on the patient's injury information. This process is time-consuming and easily affected by human factors, such as the doctor's experience and subjective judgment. Especially in emergency situations, rapid and accurate ISS scoring is particularly important.

[0003] The existing technology mainly has the following defects and shortcomings: First, the traditional ISS scoring method relies on manual operation, which is inefficient and consumes a lot of time and human resources, especially when a large number of patients need to be evaluated.

[0004] Secondly, because the scoring process relies on the subjective judgment of the doctor, different doctors may have different scoring results for the same patient, resulting in poor consistency and reliability of the evaluation results.

[0005] Finally, existing trauma information systems generally lack automated support for ISS scores, requiring doctors to enter and calculate them manually, which increases the workload and is prone to human errors. Summary of the invention

[0006] The embodiments of the present invention provide a trauma severity assessment method and system based on artificial intelligence and big data analysis, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention, Provides trauma severity assessment methods based on artificial intelligence and big data analysis, including: Extracting structured medical record data of trauma patients from a trauma treatment information system, wherein the structured medical record data includes current medical history, physical examination, auxiliary examination and diagnosis information; guiding an artificial intelligence model to identify diagnostic information related to trauma in the structured medical record data through prompt word engineering, and extracting detailed descriptions related to the diagnostic information from the structured medical record data, and generating JSON format data including diagnostic information, injury identification and injury description; Based on the split scoring strategy, each injury is scored separately, specifically including: using vector retrieval technology to retrieve scoring rules related to a single injury from the adapted ISS scoring rule knowledge base, and the ISS scoring rule knowledge base uses markdown language to adapt the original ISS scoring rules; the artificial intelligence model is based on the diagnostic information, injury identification and injury description in the JSON format data, combined with the retrieved scoring rules, to identify the area and part to which each injury belongs, and determine the corresponding injury scoring result; the injury scoring result is verified, by comparing whether the injury name in the injury scoring result is consistent with the input diagnostic information, and filtering out inconsistent scoring results; The verified injury score results are merged, and the highest score of multiple injury scores in the same part is taken as the score of the part, and the highest score of multiple part scores in the same area is taken as the score of the area; determine whether the score of each area is 6 points. If so, the total ISS score is determined to be 75 points. If not, select the three highest area scores, and take the sum of the squares of these three area scores as the total ISS score; output the total ISS score and the specific score results of each trauma area.

[0008] The artificial intelligence model is guided by the prompt word engineering to identify the diagnostic information related to the trauma in the structured medical record data, and the detailed description related to the diagnostic information is extracted from the structured medical record data, and the JSON format data containing the diagnostic information, injury identification and injury description is generated, including: A prompt word template is constructed based on a four-layer structure of role-goal-process-format, wherein the role in the prompt word template is set as an emergency doctor, the goal in the prompt word template is set as identifying trauma-related diagnoses and extracting detailed description information, the process in the prompt word template is set as diagnosis extraction, relevance judgment and description extraction, and the format in the prompt word template is set as JSON format; A multidimensional analysis method based on clinical characteristics is used to judge the relevance of trauma, which includes analyzing the time characteristics of diagnosis, judging whether it is caused by external force according to the injury mechanism, judging whether it meets the characteristics of traumatic injury in combination with the anatomical location and clinical manifestations, and establishing a trauma feature word library for feature word matching; Detailed descriptions related to diagnosis were extracted based on the contextual association analysis method, by locating the occurrence of injuries in the current medical history, extracting injury signs from physical examination records, and extracting imaging manifestations and laboratory test results from auxiliary examination results; a hierarchical extraction strategy was used to extract directly related descriptive information, indirectly related clinical manifestations, and objective examination results in turn; The extracted information is converted according to a three-layer JSON template, which includes: a diagnosis information layer for storing standardized diagnosis names, an injury identification layer for indicating the relevance of the diagnosis to the current trauma, and an injury description layer for organizing detailed information according to the dimensions of current medical history, physical examination, and auxiliary examinations; a medical terminology standardization mapping table is established to convert clinical expressions into standard medical terms, and an information integrity check mechanism is used to ensure that each diagnosis has the necessary descriptive information.

[0009] The scoring rules related to a single injury are retrieved from the adapted ISS scoring rule knowledge base using vector retrieval technology, including: The ISS scoring rules are structured and adapted, and the ISS scoring rules are classified according to the head and neck, face, chest, abdomen, limbs and body surface areas. A scoring hierarchy is established in each area according to the degree of injury. Each scoring rule is described in a standardized manner, and the standardized description includes key anatomical structures, injury types, and injury degrees, and is supplemented with common clinical variant descriptions and clinical equivalent descriptions to form a scoring rule knowledge base; The scoring rules in the scoring rule knowledge base are vectorized by using a pre-trained language model in the medical field, and each scoring rule is decomposed into three semantic segments: anatomical location, injury type, and severity. The three segments are encoded and fused to generate a rule vector representation. The rules in the same anatomical region with an injury degree less than a preset injury threshold are used as positive sample pairs, and the rules in different regions or with an injury degree greater than the preset injury threshold are used as negative sample pairs through a comparative learning mechanism, so as to minimize the vector distance between the positive sample pairs and maximize the vector distance between the negative sample pairs. A multi-level vector retrieval mechanism was constructed. Coarse-grained filtering based on anatomical regions was performed to narrow the retrieval scope. An approximate nearest neighbor retrieval index was constructed using a vector indexing algorithm. Cosine similarity was used for precise matching measurement within the candidate set. The similarity threshold was adaptively adjusted according to the rule complexity of different anatomical regions. The matching results are sorted based on vector similarity, and specific rules are given priority considering rule specificity. A rule credibility scoring mechanism is established in combination with clinical verification data to screen the most suitable scoring rules. When it is impossible to directly match a completely compliant rule, scoring recommendations are derived by analyzing the scoring patterns of similar rules and combining the principles for judging the severity of the injury.

[0010] Construct a multi-level vector retrieval mechanism, perform coarse-grained filtering based on anatomical regions to narrow the retrieval scope, use vector indexing algorithm to construct an approximate nearest neighbor retrieval index, use cosine similarity for precise matching measurement in the candidate set, and adaptively adjust the similarity threshold according to the rule complexity of different anatomical regions, including: An anatomical region index table is established to divide the human body into the head and neck, face, chest, abdomen, limbs and body surface regions, anatomical structure sub-regions are set in the anatomical region index table, scoring rules are mapped to corresponding anatomical regions through an inverted index, and the relevant regions are located in the anatomical region index table based on the anatomical position information in the input injury to obtain a first candidate rule set; A hierarchical graph structure is used to construct a vector index for the first candidate rule set. The total number of rules in the first candidate rule set is recorded as N, and the number of hierarchical layers is set to log(N). An approximate nearest neighbor graph is constructed in each layer as a sparse subset of the lower layer graph. The maximum out-degree of each layer node is set to 16 to control the density of the graph. After randomly selecting an entry point at the top layer for each rule vector in the first candidate rule set, a greedy search is performed on each layer to find the most similar node to establish a connection to form an index structure. The second candidate rule set is retrieved based on the index structure. The similarity between the rule vectors in the second candidate rule set and the input injury query vector is calculated, and the cosine similarity is used as the basic measurement indicator to calculate the similarity between the rule vector and the query vector. The position weight factor is set according to the importance of the anatomical structure, and the weight of organ injury is set higher than the weight of soft tissue injury. The weighted similarity score is calculated as a quantitative indicator of rule matching to obtain the rule set to be matched, and the similarity threshold is adaptively adjusted according to the complexity of the rules in different anatomical regions.

[0011] The artificial intelligence model identifies the region and part to which each injury belongs based on the diagnostic information, injury identification and injury description in the JSON format data, combined with the retrieved scoring rules, and determines the corresponding injury scoring results including: Pre-training a medical text corpus to obtain a pre-trained model, wherein the medical text corpus includes diagnostic reports, surgical records, and medical record writing specifications, and using a masked language model task to focus on training the semantic representation of anatomical site words and injury description words, setting the pre-trained model to a dual-stream attention structure, wherein the first attention stream is used to analyze text context semantic relations, and the second attention stream is used to identify medical entity associations, and the pre-trained model outputs a medical concept feature vector; Based on the medical concept feature vector, an anatomical knowledge graph layer is constructed, wherein the anatomical knowledge graph layer includes an anatomical knowledge base and a relational reasoning network, wherein the anatomical knowledge base uses a tree structure to store standardized human anatomical terms, and the relational reasoning network uses a graph attention network to implement message transmission, and the medical concept feature vector is mapped and aligned with the anatomical terminology to generate an anatomical feature mapping result; A hierarchical anatomical structure classifier is constructed according to the anatomical feature mapping results, the anatomical feature mapping results are divided into anatomical regions using a sequence annotation model based on a bidirectional long short-term memory network, the anatomical entity boundaries are identified using a biological information annotation scheme to obtain anatomical region annotation results, a fine-grained classification model is constructed for the anatomical region annotation results, and the features of the anatomical region annotation results are transferred to the fine-grained classification model through a residual connection mechanism to obtain an anatomical structure classification result; A semantic similarity calculation model is constructed based on the anatomical structure classification result. The semantic similarity calculation model uses a Transformer structure to vectorize the injury description and the standard description of the rule library. A correlation model between feature words and scoring results is established by analyzing historical scoring data. The gradient descent method is used to optimize the feature word weight coefficient to obtain the feature word weight vector. The feature word weight vector and the candidate rules are constructed into a directed graph structure. The rule combination pattern is learned through graph convolution operation to obtain the rule matching result. The attention pooling mechanism is used to aggregate the multiple trauma scores of the rule matching results, and the final injury score result is output.

[0012] By analyzing the historical scoring data, a correlation model between feature words and scoring results is established. The feature word weight coefficient is optimized by the gradient descent method to obtain a feature word weight vector. The feature word weight vector and the candidate rules are constructed into a directed graph structure. The rule combination pattern is learned through graph convolution operation to obtain the rule matching results including: Construct a feature word score mapping matrix, the feature word score mapping matrix includes feature word occurrence frequency vectors and corresponding score values, adopt a stratified sampling method to obtain sample sets with different score intervals, calculate the discrimination of feature words in high-score samples and low-score samples based on the sample sets to obtain initial correlation weights, and generate sample subsets through a bootstrap resampling method to calculate the average value of correlation coefficients; A feature word co-occurrence network is constructed based on the average value of the correlation coefficient, the network centrality of the feature word co-occurrence network is calculated, the network centrality is combined with the average value of the correlation coefficient to obtain a feature word weight vector, the feature word weight vector is iteratively optimized using a stochastic gradient descent method with a momentum term, the mean square error between the predicted score value and the true score value is calculated as a loss function in each iteration, and an L2 regularization term is introduced to avoid overfitting; A rule matching directed graph is constructed using the feature word weight vector, wherein the vertex set of the rule matching directed graph includes candidate rule nodes and feature word nodes, the edge set of the rule matching directed graph represents the association relationship between the rule and the feature word, and a multi-head attention mechanism is used to calculate the edge weight between the rule node and the feature word node, wherein the edge weight is jointly determined by the query vector of the rule node and the key vector and value vector of the feature word node; A graph convolutional network is applied to the rule matching directed graph, and layer-by-layer abstraction of rule node representation is achieved by stacking multiple layers of graph convolution. Each layer of graph convolution aggregates and updates the neighbor information of the node through a learnable weight matrix to obtain a high-level representation of the rule node. After the last layer of graph convolution, an attention readout mechanism is introduced to calculate the contribution of different rule nodes, and a weighted voting method is used to combine rule nodes with higher contributions to obtain the rule matching result.

[0013] A second aspect of an embodiment of the present invention provides a trauma severity assessment system based on artificial intelligence and big data analysis, comprising: The first unit is used to extract structured medical record data of trauma patients from the trauma treatment information system, wherein the structured medical record data includes current medical history, physical examination, auxiliary examination and diagnosis information; guide the artificial intelligence model to identify the diagnosis information related to trauma in the structured medical record data through prompt word engineering, and extract detailed descriptions related to the diagnosis information from the structured medical record data, and generate JSON format data including diagnosis information, injury identification and injury description; The second unit is used to score each injury separately based on the split scoring strategy, specifically including: using vector retrieval technology to retrieve scoring rules related to a single injury from the adapted ISS scoring rule knowledge base, wherein the ISS scoring rule knowledge base uses markdown language to adapt the original ISS scoring rules; the artificial intelligence model identifies the area and part to which each injury belongs based on the diagnostic information, injury identification and injury description in the JSON format data, combined with the retrieved scoring rules, and determines the corresponding injury scoring results; the injury scoring results are verified, and inconsistent scoring results are filtered out by comparing whether the injury name in the injury scoring results is consistent with the input diagnostic information; The third unit is used to merge the verified injury score results, take the highest score of multiple injury scores in the same part as the score of the part, and take the highest score of multiple part scores in the same area as the score of the area; determine whether the score of each area is 6 points. If so, the total ISS score is determined to be 75 points. If not, the three highest area scores are selected, and the sum of the squares of the three area scores is used as the total ISS score; output the total ISS score and the specific score results of each trauma area.

[0014] A third aspect of the embodiments of the present invention An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0016] The beneficial effects of this application are as follows: Automated scoring: Use AI big models to automatically extract information from medical record data and perform ISS scoring, reducing manual operations and improving scoring efficiency.

[0017] Improve scoring accuracy: Through prompt word engineering and vector retrieval technology, combined with the improved ISS scoring rule knowledge base, injuries can be identified and scored more accurately, reducing human errors and improving scoring accuracy.

[0018] Provide more detailed scoring results: not only the total ISS score is provided, but also the specific scoring results of each trauma area are provided, providing clinicians with more comprehensive information and assisting in formulating more accurate treatment plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of a flow chart of a method for assessing the severity of trauma using artificial intelligence and big data analysis according to an embodiment of the present invention; Figure 2 Schematic diagram of the structure of the trauma severity assessment system based on artificial intelligence and big data analysis according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0022] Figure 1 FIG. 1 is a flow chart of a method for assessing the severity of trauma using artificial intelligence and big data analysis according to an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Extract structured medical record data of trauma patients from a trauma treatment information system, wherein the structured medical record data includes current medical history, physical examination, auxiliary examination and diagnosis information; guide the artificial intelligence model to identify diagnostic information related to trauma in the structured medical record data through prompt word engineering, and extract detailed descriptions related to the diagnostic information from the structured medical record data, and generate JSON format data including diagnostic information, injury identification and injury description; S102. Based on the split scoring strategy, each injury is scored separately, specifically including: using vector retrieval technology to retrieve scoring rules related to a single injury from the adapted ISS scoring rule knowledge base, wherein the ISS scoring rule knowledge base uses markdown language to adapt the original ISS scoring rules; the artificial intelligence model identifies the region and part to which each injury belongs based on the diagnostic information, injury identification and injury description in the JSON format data, combined with the retrieved scoring rules, and determines the corresponding injury scoring result; the injury scoring result is verified, and the inconsistent scoring results are filtered out by comparing whether the injury name in the injury scoring result is consistent with the input diagnostic information; S103. Merge the verified injury score results, take the highest score among multiple injury scores in the same part as the score of the part, and take the highest score among multiple part scores in the same area as the score of the area; determine whether the score of each area is 6 points. If so, determine the total ISS score to be 75 points. If not, select the three highest area scores, and take the sum of the squares of the three area scores as the total ISS score; output the total ISS score and the specific score results of each trauma area.

[0023] For example, case data extraction: The medical records of trauma patients have been structured from the "National Severe Trauma Treatment Information Interactive Linkage System" or the "Trauma Center Business Management System" (the medical records are parsed into structured fields such as chief complaint, current medical history, physical examination, and auxiliary examinations). When performing ISS automatic scoring, the relevant fields of ISS scoring are automatically extracted from the system, including current medical history, physical examination, auxiliary examination, and diagnosis. Combined with the prompt word project, the AI ​​big model semantically understands the key information of the medical record, automatically identifies the diagnosis related to this trauma, and extracts detailed medical record content to generate the corresponding injury description.

[0024] Guide the AI ​​big model to solve the ISS scoring problem through prompt word engineering.

[0025] 1) Set the role of the big model as an emergency doctor who is good at analyzing medical records and identifying injury information related to trauma in the medical records; 2) Give the big model processing goals; 3) Prompt the big model processing process: a. Extract the diagnosis in the input text; b. Identify whether each diagnosis is related to this acute trauma; c. For the diagnosis related to this trauma, extract detailed descriptions related to the diagnosis from the medical records (medical history, physical examination, auxiliary examinations, etc.); 4) Specify the output format and output "diagnosis", "whether there is injury", and "injury description" in JSON format.

[0026] A split scoring strategy is adopted, based on the medical knowledge base adapted from the ISS scoring rules (2005 international version). The AI ​​big model combines prompt word engineering to conduct in-depth analysis of each injury, identify the area and part to which each injury belongs, and give a corresponding injury score.

[0027] Explanation of the split scoring strategy: When the AI ​​large model performs ISS scoring, it needs to retrieve the scoring rules related to the injury in the knowledge base in order to accurately score. When scoring multiple injuries, the multiple injuries of multiple injuries may have very different semantics and are relatively discrete in the vector database, resulting in a low correlation between the retrieved scoring rules and the injury, and the accuracy of the final scoring result will be very low. This method splits the scoring process, first identifying the injury, then scoring each injury one by one, and finally integrating the scores of each injury to obtain the final ISS score. This strategy improves the knowledge base retrieval of only a single semantic injury in each injury scoring, ensures the relevance of the retrieval results to the injury, and ultimately ensures the accuracy of the scoring.

[0028] According to the injuries identified in the previous step, each injury is scored.

[0029] The ISS scoring rules were adapted. The original ISS scoring rules were a table with 9 columns and 50 rows. The content was also extremely simplified and summarized. It was difficult for the AI ​​model to have a good semantic understanding of the original ISS scoring rules. For example, the 2-point scoring rule for "Region 5, Limbs, Pelvis, and Hips" - "Bone" is: "Limb fractures (excluding 1 / 3 / 4 / 5-point fractures) [2-point details: The following fractures "clavicle; scapula; humerus / radius / ulna (excluding: open non-simple diaphyseal fractures, open fractures involving joints, but ulna styloid process open fracture is 2 points); carpal bone; metacarpal bone; tibia closed; fibula; patella; talus; calcaneus; cuneiform bone; cuboid bone; metatarsal bone; navicular bone; pelvic ring fracture NFS; pelvic ring stable fracture with intact posterior ring; closed acetabular fracture"]; pelvic fracture NFS / stable fracture with intact posterior ring; acetabular fracture", which mixes multiple parts, injury conditions and injury degrees together. It is difficult for the AI ​​large model to recognize. Through adaptation, the table is rewritten using markdown language and the mixed description is expanded.

[0030] The scoring results were checked and those that did not meet the diagnosis were eliminated to avoid the influence of large model illusion and ensure the accuracy of the scoring.

[0031] Verify whether the "injury" name in the injury scoring result is consistent with the "diagnosis" name entered for scoring, and filter out those that are inconsistent.

[0032] The AI ​​big model is a generative big model, and there is an illusion that the generated results do not match the requirements. For example, if the injury being scored is A, the scoring result may output multiple injury results: injury scoring result A and injury scoring result B. The injury scoring result B is not needed. At this time, it is necessary to check and filter out this part of the injury to ensure the accuracy of subsequent integration.

[0033] Score merging: For multiple injuries in the same trauma area or part, the system automatically merges the injuries in the same area and part to obtain the regional score. When merging injuries in the same part, if there are multiple injuries in the same part, the highest injury score is taken as the score for that part. When merging injuries in the same area, if there are multiple injuries in the same area, the highest site score is taken as the score for that area.

[0034] ISS total score calculation and output: Calculate the ISS total score based on the regional scoring results, and output the specific scores for each region and part. Determine whether the regional score is 6 points. If so, the ISS total score is 75 points; if not, continue to calculate. According to the regional scores from high to low, take the results with the highest scores in the first three regions. ISS score = the sum of the squares of the scores of the first three regions. Scoring result output: The system finally outputs the ISS total score and the specific scores of each trauma area for reference by clinicians and to assist in decision-making.

[0035] The automatic scoring system of the present invention can effectively improve the efficiency and accuracy of trauma assessment and reduce human errors, especially in the treatment of multiple injuries and complex injuries. This method not only provides a fast and accurate auxiliary decision-making tool for medical personnel in the fields of emergency and trauma, but also improves the timeliness and accuracy of trauma treatment in clinical applications, providing an intelligent solution for trauma management.

[0036] The present invention has broad application prospects and can significantly improve the automation level of trauma scoring in the medical field, reduce the workload of doctors, optimize the trauma management process, and improve the treatment effect of patients.

[0037] In an optional implementation, the prompt word engineering is used to guide the artificial intelligence model to identify the diagnostic information related to the trauma in the structured medical record data, and to extract the detailed description related to the diagnostic information from the structured medical record data, and to generate JSON format data containing the diagnostic information, the injury identification and the injury description, including: A prompt word template is constructed based on a four-layer structure of role-goal-process-format, wherein the role in the prompt word template is set as an emergency doctor, the goal in the prompt word template is set as identifying trauma-related diagnoses and extracting detailed description information, the process in the prompt word template is set as diagnosis extraction, relevance judgment and description extraction, and the format in the prompt word template is set as JSON format; A multidimensional analysis method based on clinical characteristics is used to judge the relevance of trauma, which includes analyzing the time characteristics of diagnosis, judging whether it is caused by external force according to the injury mechanism, judging whether it meets the characteristics of traumatic injury in combination with the anatomical location and clinical manifestations, and establishing a trauma feature word library for feature word matching; Detailed descriptions related to diagnosis were extracted based on the contextual association analysis method, by locating the occurrence of injuries in the current medical history, extracting injury signs from physical examination records, and extracting imaging manifestations and laboratory test results from auxiliary examination results; a hierarchical extraction strategy was used to extract directly related descriptive information, indirectly related clinical manifestations, and objective examination results in turn; The extracted information is converted according to a three-layer JSON template, which includes: a diagnosis information layer for storing standardized diagnosis names, an injury identification layer for indicating the relevance of the diagnosis to the current trauma, and an injury description layer for organizing detailed information according to the dimensions of current medical history, physical examination, and auxiliary examinations; a medical terminology standardization mapping table is established to convert clinical expressions into standard medical terms, and an information integrity check mechanism is used to ensure that each diagnosis has the necessary descriptive information.

[0038] First, a prompt word template based on the four-layer structure of role-goal-process-format was constructed. The role was set as "emergency doctor", the goal was set as "identify trauma-related diagnoses and extract detailed description information", the process was set as "diagnosis extraction, relevance judgment and description extraction", and the format was set as "JSON".

[0039] For example, the input prompt word can be: "You are an emergency doctor who needs to identify all trauma-related diagnoses from the patient's structured medical record data, extract detailed description information, and finally output the results in JSON format."

[0040] Next, perform diagnosis extraction. Extract all diagnosis information from the diagnosis list of structured medical record data. For example, extract diagnosis information such as "S06.001 Scalp contusion", "S52.502a Open fracture of right forearm", and "J18.9 Pneumonia, unspecified" from the medical record data.

[0041] Then, the trauma relevance is judged. A multidimensional analysis method based on clinical characteristics is used, including: analyzing the time characteristics of the diagnosis, for example, if the diagnosis time is close to the admission time, it indicates that it may be related to the current trauma; judging whether it is caused by external force according to the injury mechanism, for example, descriptions such as "fall", "collision", "traffic accident" indicate external force injury; combining anatomical location and clinical manifestations to judge whether it meets the characteristics of traumatic injury, such as "pain", "swelling", "limited activity", etc. At the same time, a trauma feature word library is established, such as words including "fracture", "contusion", "laceration", "open wound", etc., for feature word matching. For example, "S06.001 scalp contusion" and "S52.502a right forearm open fracture" meet the characteristics of traumatic injury, while "J18.9 pneumonia, unspecified" does not.

[0042] After that, descriptive information extraction is performed. Based on the contextual association analysis method, detailed descriptions related to diagnosis are extracted from different parts of the medical record data. The occurrence of injury is located in the current medical history, such as the patient was admitted to the hospital due to "head trauma caused by a car accident"; the signs of injury are extracted from the physical examination records, such as "right forearm deformity, swelling, and tenderness"; and the imaging manifestations and laboratory results are extracted from the auxiliary examination results, such as "right forearm X-ray shows ulna and radius fractures." A hierarchical extraction strategy is adopted to extract directly related descriptive information, indirectly related clinical manifestations, and objective examination results in turn. For example, for "S52.502a right forearm open fracture", "right forearm deformity, swelling, tenderness" and "right forearm X-ray shows ulna and radius fractures" can be directly extracted as directly related descriptive information, and the vital signs of the patient at admission can be extracted as indirectly related clinical manifestations.

[0043] Finally, the extracted information is converted into JSON format for output. A three-layer structure is used: the diagnostic information layer stores standardized diagnostic names, such as converting "S52.502a Open fracture of right forearm" into ICD-10 standard codes; the injury identification layer indicates the relevance of the diagnosis to the current trauma, such as "yes" or "no"; the injury description layer organizes detailed information according to the dimensions of current medical history, physical examination, and auxiliary examination.

[0044] Establish a standardized mapping table for medical terms, such as converting "scalp laceration" to "S01.001" to ensure the standardization of diagnosis names. Establish an information integrity check mechanism to ensure that each diagnosis has the necessary descriptive information. For example, if a diagnosis lacks a description of the current medical history, provide a prompt or supplement.

[0045] The beneficial effects of this technical solution can be summarized into the following three aspects: Improve the accuracy and efficiency of trauma-related information extraction: Through multi-dimensional analysis and contextual correlation analysis, it is possible to more accurately identify trauma-related diagnoses and extract more complete descriptive information, reducing the workload of manual review and improving efficiency.

[0046] Facilitates the standardized and structured storage of trauma data: The extracted information is converted into a standard JSON format to facilitate data storage, retrieval, and analysis, providing a basis for subsequent trauma research and statistical analysis.

[0047] Support trauma diagnosis and treatment decisions: The extracted structured trauma information can provide doctors with more comprehensive information about the condition, assisting doctors in making more accurate diagnoses and formulating more reasonable treatment plans.

[0048] In an optional implementation, using vector retrieval technology to retrieve scoring rules related to a single injury from the adapted ISS scoring rule knowledge base includes: The ISS scoring rules are structured and adapted, and the ISS scoring rules are classified according to the head and neck, face, chest, abdomen, limbs and body surface areas. A scoring hierarchy is established in each area according to the degree of injury. Each scoring rule is described in a standardized manner, and the standardized description includes key anatomical structures, injury types, and injury degrees, and is supplemented with common clinical variant descriptions and clinical equivalent descriptions to form a scoring rule knowledge base; The scoring rules in the scoring rule knowledge base are vectorized by using a pre-trained language model in the medical field, and each scoring rule is decomposed into three semantic segments: anatomical location, injury type, and severity. The three segments are encoded and fused to generate a rule vector representation. The rules in the same anatomical region with an injury degree less than a preset injury threshold are used as positive sample pairs, and the rules in different regions or with an injury degree greater than the preset injury threshold are used as negative sample pairs through a comparative learning mechanism, so as to minimize the vector distance between the positive sample pairs and maximize the vector distance between the negative sample pairs. A multi-level vector retrieval mechanism was constructed. Coarse-grained filtering based on anatomical regions was performed to narrow the retrieval scope. An approximate nearest neighbor retrieval index was constructed using a vector indexing algorithm. Cosine similarity was used for precise matching measurement within the candidate set. The similarity threshold was adaptively adjusted according to the rule complexity of different anatomical regions. The matching results are sorted based on vector similarity, and specific rules are given priority considering rule specificity. A rule credibility scoring mechanism is established in combination with clinical verification data to screen the most suitable scoring rules. When it is impossible to directly match a completely compliant rule, scoring recommendations are derived by analyzing the scoring patterns of similar rules and combining the principles for judging the severity of the injury.

[0049] The scoring rules related to a single injury condition are retrieved from the adapted ISS scoring rule knowledge base. The specific implementation method is as follows: First, the ISS scoring rules were structured and adapted to build a scoring rule knowledge base. The ISS scoring rules were classified according to human anatomical regions, such as head and neck, face, chest, abdomen, limbs, and body surface. In each region, a scoring hierarchy was established according to the severity of the injury, such as mild, moderate, and severe. Each scoring rule was described in a standardized manner, including key anatomical structures, injury types, and injury severity. For example, for "left femoral shaft fracture", the key anatomical structure is "left femoral shaft", the injury type is "fracture", and the injury severity can be judged based on the specific circumstances of the fracture. In order to improve the accuracy of the retrieval, common clinical variant expressions and clinical equivalent descriptions were supplemented, such as "left femoral fracture" can be used as a variant expression of "left femoral shaft fracture". It is assumed that the scoring rules for the head and neck region include: scalp laceration (mild), skull fracture (moderate), brain contusion (severe), etc.

[0050] Then, each rule in the scoring rule knowledge base is vectorized using a pre-trained language model in the medical field. Each rule is decomposed into three semantic segments: anatomical location, injury type, and severity for encoding. For example, for "left femoral shaft fracture", the anatomical location is encoded as a vector representation of "left femoral shaft", the injury type is encoded as a vector representation of "fracture", and the severity is encoded as a vector representation of the specific degree. The vector representations of the three semantic segments are fused to generate the vector representation of the rule. In order to improve the quality of the vector representation, a contrastive learning mechanism is adopted. Rules with the same anatomical region and an injury degree less than a preset threshold are taken as positive sample pairs, such as "left femoral shaft crack fracture" and "left femoral shaft greenstick fracture". Rules with different regions or an injury degree greater than a preset threshold are taken as negative sample pairs, such as "left femoral shaft fracture" and "right humeral fracture", or "left femoral shaft comminuted fracture" and "left femoral shaft crack fracture" (assuming that the severity of comminuted fracture exceeds the preset threshold). Through contrastive learning, the vector distance between positive sample pairs is minimized and the vector distance between negative sample pairs is maximized.

[0051] Next, a multi-level vector retrieval mechanism is constructed. First, coarse-grained filtering is performed based on anatomical regions to narrow the search scope. For example, if the input injury description is "left femoral shaft fracture", only the scoring rules of the "limbs" region are retrieved. Then, the vector index algorithm is used to construct an approximate nearest neighbor retrieval index. Within the candidate set, cosine similarity is used for exact matching measurement. The similarity threshold is adaptively adjusted according to the complexity of the rules in different anatomical regions. For example, the rules in the head and neck region are more complex, and the similarity threshold can be set higher, while the rules in the body surface region are simpler, and the similarity threshold can be set lower.

[0052] Finally, the matching results are sorted based on vector similarity. Rules with specific descriptions are preferred. For example, "comminuted fracture of left femoral shaft" is more specific than "fracture of left femoral shaft". A rule credibility scoring mechanism is established in combination with clinical validation data to screen the most suitable scoring rules. When it is impossible to directly match a completely matching rule, the scoring suggestions are derived by analyzing the scoring patterns of similar rules and combining the principles for judging the severity of the injury. For example, if a rule that completely matches "fracture of left femoral neck with femoral head necrosis" cannot be found, the scoring rules of "fracture of left femoral neck" and "femoral head necrosis" can be referred to and derived in combination with the principles for judging the severity of the injury.

[0053] The beneficial effects of this method can be summarized in the following three aspects: 1. Improve retrieval efficiency: Multi-level retrieval mechanism and vector index algorithm can quickly screen candidate rules and shorten retrieval time.

[0054] 2. Improve retrieval accuracy: Semantic fragment encoding, contrastive learning, and adaptive similarity thresholds can improve matching accuracy.

[0055] 3. Provide scoring suggestions: Even if an exact match is not possible, scoring suggestions can be provided based on similar rules and damage levels to enhance the practicality of the system.

[0056] In an optional implementation, a multi-level vector retrieval mechanism is constructed, a coarse-grained filtering is performed based on the anatomical region to narrow the retrieval range, an approximate nearest neighbor retrieval index is constructed using a vector indexing algorithm, and a cosine similarity is used in the candidate set for accurate matching measurement. The similarity threshold is adaptively adjusted according to the rule complexity of different anatomical regions, including: An anatomical region index table is established to divide the human body into the head and neck, face, chest, abdomen, limbs and body surface regions, anatomical structure sub-regions are set in the anatomical region index table, scoring rules are mapped to corresponding anatomical regions through an inverted index, and the relevant regions are located in the anatomical region index table based on the anatomical position information in the input injury to obtain a first candidate rule set; A hierarchical graph structure is used to construct a vector index for the first candidate rule set. The total number of rules in the first candidate rule set is recorded as N, and the number of hierarchical layers is set to log(N). An approximate nearest neighbor graph is constructed in each layer as a sparse subset of the lower layer graph. The maximum out-degree of each layer node is set to 16 to control the density of the graph. After randomly selecting an entry point at the top layer for each rule vector in the first candidate rule set, a greedy search is performed on each layer to find the most similar node to establish a connection to form an index structure. The second candidate rule set is retrieved based on the index structure. The similarity between the rule vectors in the second candidate rule set and the input injury query vector is calculated, and the cosine similarity is used as the basic measurement indicator to calculate the similarity between the rule vector and the query vector. The position weight factor is set according to the importance of the anatomical structure, and the weight of organ injury is set higher than the weight of soft tissue injury. The weighted similarity score is calculated as a quantitative indicator of rule matching to obtain the rule set to be matched, and the similarity threshold is adaptively adjusted according to the complexity of the rules in different anatomical regions.

[0057] The multi-level vector retrieval injury rule matching method is used to quickly and accurately match the corresponding scoring rules according to the injury description.

[0058] First, construct an anatomical region index table. Divide the human body into more detailed anatomical regions, for example, the head and neck are subdivided into blood vessels, nerves, bones, etc.; the chest is subdivided into diaphragm, esophagus, lungs, heart, etc.; the abdomen is subdivided into hollow organs, important solid organs, lumbar vertebrae, etc.; the limbs are subdivided into bones, joints, tendons and ligaments, etc.; the body surface area is subdivided into skin, subcutaneous tissue, etc. Each anatomical region has multiple sub-regions. For example, the face can be subdivided into eyes, nose, mouth, etc., and important solid organs can be subdivided into liver, spleen, kidney, etc. Then, associate the scoring rules with the corresponding anatomical regions and sub-regions, and achieve fast retrieval through inverted indexing. For example, the "traumatic spleen rupture" rule is associated with "abdomen-important solid organs-spleen", and the "anterior cruciate ligament rupture" rule is associated with "limbs-tendon ligament-cruciate ligament".

[0059] After the patient's medical information is entered, the anatomical location information in the injury description is extracted. For example, if the injury description is "the patient has head trauma, accompanied by left orbital fracture", the two anatomical location information of "head" and "left orbit" are extracted. This information is used to locate in the anatomical region index table, for example, "head" is located in the "head and neck" region, and "left orbit" is located in the "head and neck-face-left orbit" sub-region. All scoring rules associated with these regions are obtained to form the first candidate rule set. In this example, it is assumed that the first candidate rule set contains rules such as "open craniocerebral injury", "closed craniocerebral injury", "orbital fracture", and "facial soft tissue injury".

[0060] Next, construct a hierarchical graph vector index for the first candidate rule set. Assume that the first candidate rule set contains 1024 rules. Set the number of hierarchical layers to log 21024 = 10. Each layer builds an approximate nearest neighbor graph as a sparse subset of the lower layer graph. The maximum out-degree of each layer node is set to 16 to control the density of the graph. Each rule is vectorized, for example, the rule text is converted into a vector using a word vector model. Then, a random entry point is selected from the top layer, and a greedy search is performed on each layer to find the node most similar to the current rule vector and establish a connection. Finally, a ten-layer index structure is formed.

[0061] Retrieve the second candidate rule set based on the constructed index structure. Vectorize the injury query, for example, convert "the patient has head trauma with left orbital fracture" into a vector. Start from the top entry point, search down layer by layer, find the node most similar to the query vector at each layer, and finally reach the bottom layer to obtain the rules associated with these nodes to form the second candidate rule set. Assume that the second candidate rule set in this example contains rules such as "open craniocerebral injury", "orbital fracture", and "facial soft tissue injury".

[0062] Finally, the second candidate rule set is matched exactly. The cosine similarity between each rule vector in the second candidate rule set and the injury query vector is calculated. The position weight factor is set according to the importance of the anatomical structure. For example, the weight of "brain" is set to 0.9, the weight of "orbit" is set to 0.8, and the weight of "facial soft tissue" is set to 0.7. The position weight factor is multiplied by the cosine similarity to obtain a weighted similarity score. Assuming that the cosine similarity of "open craniocerebral injury" is 0.85, its weighted similarity score is 0.85 * 0.9 = 0.765. The similarity threshold is adaptively adjusted according to the complexity of the rules in different anatomical regions. For example, the threshold of the "brain" region is set to 0.8, and the threshold of the "face" region is set to 0.7. Compare the weighted similarity score with the threshold to filter out the final matching rules. Assume that the final matching rules in this example are "open craniocerebral injury" and "orbital fracture".

[0063] The beneficial effects of this method are reflected in three aspects: 1. Improved retrieval efficiency: Through multi-level vector indexing and coarse-grained anatomical region filtering, the retrieval scope is significantly narrowed and the retrieval efficiency is improved.

[0064] 2. Improved matching accuracy: The matching accuracy is improved by combining vector similarity calculation and a weighting mechanism based on the importance of anatomical regions, as well as adaptively adjusting the similarity threshold according to the complexity of the rules.

[0065] 3. Enhanced system maintainability: The structured design of the anatomical region index table and the association between rules and anatomical regions make it easier to maintain and update the rule base.

[0066] In an optional implementation, the artificial intelligence model identifies the region and part to which each injury belongs based on the diagnostic information, injury identification and injury description in the JSON format data, combined with the retrieved scoring rules, and determines the corresponding injury scoring results including: Pre-training a medical text corpus to obtain a pre-trained model, wherein the medical text corpus includes diagnostic reports, surgical records, and medical record writing specifications, and using a masked language model task to focus on training the semantic representation of anatomical site words and injury description words, setting the pre-trained model to a dual-stream attention structure, wherein the first attention stream is used to analyze text context semantic relations, and the second attention stream is used to identify medical entity associations, and the pre-trained model outputs a medical concept feature vector; Based on the medical concept feature vector, an anatomical knowledge graph layer is constructed, wherein the anatomical knowledge graph layer includes an anatomical knowledge base and a relational reasoning network, wherein the anatomical knowledge base uses a tree structure to store standardized human anatomical terms, and the relational reasoning network uses a graph attention network to implement message transmission, and the medical concept feature vector is mapped and aligned with the anatomical terminology to generate an anatomical feature mapping result; A hierarchical anatomical structure classifier is constructed according to the anatomical feature mapping results, the anatomical feature mapping results are divided into anatomical regions using a sequence annotation model based on a bidirectional long short-term memory network, the anatomical entity boundaries are identified using a biological information annotation scheme to obtain anatomical region annotation results, a fine-grained classification model is constructed for the anatomical region annotation results, and the features of the anatomical region annotation results are transferred to the fine-grained classification model through a residual connection mechanism to obtain an anatomical structure classification result; A semantic similarity calculation model is constructed based on the anatomical structure classification result. The semantic similarity calculation model uses a Transformer structure to vectorize the injury description and the standard description of the rule library. A correlation model between feature words and scoring results is established by analyzing historical scoring data. The gradient descent method is used to optimize the feature word weight coefficient to obtain the feature word weight vector. The feature word weight vector and the candidate rules are constructed into a directed graph structure. The rule combination pattern is learned through graph convolution operation to obtain the rule matching result. The attention pooling mechanism is used to aggregate the multiple trauma scores of the rule matching results, and the final injury score result is output.

[0067] The specific implementation method of the intelligent injury scoring method based on the injury information and scoring rules in the JSON format data is as follows: First, a medical text corpus is constructed and a pre-trained model is trained. The corpus contains a large number of diagnostic reports, surgical records, and medical record writing specifications, covering a wealth of medical knowledge. The Masked Language Model (MLM) is used as a training task, focusing on the semantic representation learning of anatomical part words and injury description words. For example, the word "humerus" in the sentence "the patient's left humerus fracture" is masked, and the model needs to predict the masked word based on the context. With a large amount of training data, the model can learn the semantic association between "humerus" and words such as "left" and "fracture", thereby obtaining a more accurate word vector representation. The pre-trained model adopts a two-stream attention structure. The first attention stream analyzes the semantic relationship of the text context. For example, in "left humerus fracture with nerve injury", it focuses on the relationship between "fracture" and "nerve injury". The second attention stream identifies medical entity associations, such as recognizing that "humerus" is an anatomical part of the human body. Finally, the pre-trained model outputs a medical concept feature vector containing rich medical semantic information.

[0068] Then, the anatomical knowledge graph layer is constructed. The graph layer contains the anatomical knowledge base and the relational reasoning network. The anatomical knowledge base uses a tree structure to store standardized human anatomical terms, for example, "human body" - "upper limb" - "arm" - "humerus". The relational reasoning network uses a graph attention network (GAT) to implement message passing. The medical concept feature vectors output by the pre-trained model are mapped and aligned with the anatomical terms in the anatomical knowledge base. For example, the "humerus" in "humeral fracture" is matched with the "humerus" in the anatomical knowledge base. Through the relational reasoning network, the association between different anatomical parts can be learned, such as "humerus" belongs to "arm", thereby generating a more comprehensive anatomical feature mapping result.

[0069] Next, a hierarchical anatomical structure classifier is constructed. First, a sequence annotation model based on a bidirectional long short-term memory network (BiLSTM) is used to divide the anatomical feature mapping results into anatomical regions. For example, the "left humerus" in "left humerus fracture" is identified as the "upper limb" region. Then, the anatomical entity boundaries are identified using a bioinformatics annotation scheme (e.g., BIO annotation) to obtain the anatomical region annotation results. For example, "left humerus fracture" is annotated as "B-upper limb I-upper limb OO". Finally, a fine-grained classification model is constructed for the anatomical region annotation results. The features of the anatomical region annotation results are transferred to the fine-grained classification model through the residual connection mechanism to obtain a more accurate anatomical structure classification result, such as further classifying "left humerus" into "humerus".

[0070] Finally, a semantic similarity calculation model is constructed and the injury score is performed. The model uses the Transformer structure to vectorize the injury description and the standard description of the rule base. For example, "left humeral fracture" is vectorized with "humeral fracture" in the rule base. A correlation model between feature words and scoring results is established by analyzing historical scoring data. For example, the analysis found that feature words such as "openness" and "comminuted" are associated with higher scoring results. The feature word weight coefficient is optimized by the gradient descent method to obtain the feature word weight vector. The feature word weight vector and the candidate rules are constructed into a directed graph structure. The rule combination pattern is learned through graph convolution operations to obtain the rule matching result. For example, "open comminuted fracture of left humerus" is matched with multiple rules. The attention pooling mechanism is used to aggregate the multiple trauma scores of the rule matching results and output the final injury score result. For example, according to the matched rules, the final score of the injury is calculated to be 3 points.

[0071] Data example: The input JSON data contains the injury identification "1" and the injury description "open comminuted fracture of humerus". Through the above steps, the final output injury identification is "1", the anatomical region is "limbs", the anatomical part is "bone", and the injury score is 3 points.

[0072] Beneficial effects: Improve scoring efficiency: Automate injury scoring, reduce manual operations, and improve scoring efficiency.

[0073] Improve scoring accuracy: Based on knowledge graphs and deep learning models, improve scoring accuracy and reduce human errors.

[0074] Improve scoring consistency: Standardize the scoring process to improve the consistency of scoring between different doctors.

[0075] In an optional implementation, a correlation model between feature words and scoring results is established by analyzing historical scoring data, a feature word weight coefficient is optimized by a gradient descent method to obtain a feature word weight vector, the feature word weight vector and the candidate rule are constructed into a directed graph structure, and a rule combination pattern is learned by a graph convolution operation to obtain a rule matching result including: Construct a feature word score mapping matrix, the feature word score mapping matrix includes feature word occurrence frequency vectors and corresponding score values, adopt a stratified sampling method to obtain sample sets with different score intervals, calculate the discrimination of feature words in high-score samples and low-score samples based on the sample sets to obtain initial correlation weights, and generate sample subsets through a bootstrap resampling method to calculate the average value of correlation coefficients; A feature word co-occurrence network is constructed based on the average value of the correlation coefficient, the network centrality of the feature word co-occurrence network is calculated, the network centrality is combined with the average value of the correlation coefficient to obtain a feature word weight vector, the feature word weight vector is iteratively optimized using a stochastic gradient descent method with a momentum term, the mean square error between the predicted score value and the true score value is calculated as a loss function in each iteration, and an L2 regularization term is introduced to avoid overfitting; A rule matching directed graph is constructed using the feature word weight vector, wherein the vertex set of the rule matching directed graph includes candidate rule nodes and feature word nodes, the edge set of the rule matching directed graph represents the association relationship between the rule and the feature word, and a multi-head attention mechanism is used to calculate the edge weight between the rule node and the feature word node, wherein the edge weight is jointly determined by the query vector of the rule node and the key vector and value vector of the feature word node; A graph convolutional network is applied to the rule matching directed graph, and layer-by-layer abstraction of rule node representation is achieved by stacking multiple layers of graph convolution. Each layer of graph convolution aggregates and updates the neighbor information of the node through a learnable weight matrix to obtain a high-level representation of the rule node. After the last layer of graph convolution, an attention readout mechanism is introduced to calculate the contribution of different rule nodes, and a weighted voting method is used to combine rule nodes with higher contributions to obtain the rule matching result.

[0076] The method for constructing and applying a correlation model between feature words and scoring results is to analyze historical scoring data, establish a correlation model between feature words and scoring results, and use a graph convolutional network to learn rule combination patterns to ultimately obtain rule matching results.

[0077] First, construct a feature word score mapping matrix. This matrix contains the feature word frequency vector and the corresponding score value. These data form the basis of the feature word score mapping matrix. In order to calculate the correlation more accurately, a stratified sampling method is used to obtain sample sets with different score intervals. For example, 1,000 injury data are divided into three intervals according to scores of 1-2 points, 3-4 points, and 4-5 points. A certain number of samples are randomly selected from each interval to ensure that the sample set can represent the overall data distribution. Then, based on these sampled sample sets, the discrimination of feature words in high-score samples and low-score samples is calculated to obtain the initial correlation weight. In order to make the results more stable, the bootstrap resampling method is used to extract subsets from the sample set with replacement multiple times, calculate the correlation coefficient each time, and finally take the average value as the final correlation coefficient.

[0078] Next, a feature word co-occurrence network is constructed based on the average value of the correlation coefficient. The nodes in the network represent feature words. If two feature words appear in the same injury data, there is an edge between the two nodes. The centrality of each node in the network is calculated, such as degree centrality, betweenness centrality, etc. The network centrality is combined with the average value of the correlation coefficient to obtain the feature word weight vector. In order to optimize the feature word weight vector, the stochastic gradient descent method with momentum term is used for iterative optimization. In each iteration, the feature word weight vector is used to predict the score, and the mean square error between the predicted score and the actual score is calculated as the loss function. At the same time, in order to avoid overfitting, the L2 regularization term is introduced to limit the size of the numerical value in the weight vector.

[0079] Then, the feature word weight vector is used to construct a rule matching directed graph. The vertex set in the graph contains candidate rule nodes and feature word nodes. The edge set represents the association between rules and feature words. The multi-head attention mechanism is used to calculate the edge weights between rule nodes and feature word nodes. Each attention head has a set of query vectors, key vectors, and value vectors. The query vector of the rule node and the key vector and value vector of the feature word node jointly determine the edge weight.

[0080] Finally, a graph convolutional network is applied to the rule matching directed graph. By stacking multiple layers of graph convolution, the layer-by-layer abstraction of the rule node representation is achieved. Each layer of graph convolution aggregates and updates the node's neighbor information through a learnable weight matrix to obtain a high-level representation of the rule node. After the last layer of graph convolution, an attention readout mechanism is introduced to calculate the contribution of different rule nodes. Finally, a weighted voting method is used to combine the rule nodes with higher contributions to obtain the rule matching result.

[0081] Beneficial effects: 1. Improve the accuracy of rating prediction: By analyzing historical rating data, extracting key feature words, and combining graph convolutional network learning rule combination patterns, ratings can be predicted more accurately, thereby improving the accuracy of rating prediction.

[0082] 2. Enhance model interpretability: This method visualizes the feature word weights and rule matching results, which can better understand the decision-making process of the model, enhance the interpretability of the model, and make it easier for users to understand and trust the model results.

[0083] 3. Improve rule matching efficiency: Graph convolutional networks can be used to efficiently learn rule combination patterns, avoid complex rule matching processes, improve rule matching efficiency, and are suitable for large-scale data processing.

[0084] Figure 2 FIG. 1 is a schematic diagram of the structure of a trauma severity assessment system based on artificial intelligence and big data analysis according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to extract structured medical record data of trauma patients from the trauma treatment information system, wherein the structured medical record data includes current medical history, physical examination, auxiliary examination and diagnosis information; guide the artificial intelligence model to identify the diagnosis information related to trauma in the structured medical record data through prompt word engineering, and extract detailed descriptions related to the diagnosis information from the structured medical record data, and generate JSON format data including diagnosis information, injury identification and injury description; The second unit is used to score each injury separately based on the split scoring strategy, specifically including: using vector retrieval technology to retrieve scoring rules related to a single injury from the adapted ISS scoring rule knowledge base, wherein the ISS scoring rule knowledge base uses markdown language to adapt the original ISS scoring rules; the artificial intelligence model identifies the area and part to which each injury belongs based on the diagnostic information, injury identification and injury description in the JSON format data, combined with the retrieved scoring rules, and determines the corresponding injury scoring results; the injury scoring results are verified, and inconsistent scoring results are filtered out by comparing whether the injury name in the injury scoring results is consistent with the input diagnostic information; The third unit is used to merge the verified injury score results, take the highest score of multiple injury scores in the same part as the score of the part, and take the highest score of multiple part scores in the same area as the score of the area; determine whether the score of each area is 6 points. If so, the total ISS score is determined to be 75 points. If not, the three highest area scores are selected, and the sum of the squares of the three area scores is used as the total ISS score; output the total ISS score and the specific score results of each trauma area.

[0085] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0086] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0087] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A trauma severity assessment method based on artificial intelligence and big data analysis, characterized in that: include: Extracting structured medical record data of trauma patients from a trauma treatment information system, wherein the structured medical record data includes current medical history, physical examination, auxiliary examination and diagnosis information; guiding an artificial intelligence model to identify diagnostic information related to trauma in the structured medical record data through prompt word engineering, and extracting detailed descriptions related to the diagnostic information from the structured medical record data, and generating JSON format data including diagnostic information, injury identification and injury description; Based on the split scoring strategy, each injury is scored separately, specifically including: using vector retrieval technology to retrieve scoring rules related to a single injury from the adapted ISS scoring rule knowledge base, and the ISS scoring rule knowledge base uses markdown language to adapt the original ISS scoring rules; the artificial intelligence model is based on the diagnostic information, injury identification and injury description in the JSON format data, combined with the retrieved scoring rules, to identify the area and part to which each injury belongs, and determine the corresponding injury scoring result; the injury scoring result is verified, by comparing whether the injury name in the injury scoring result is consistent with the input diagnostic information, and filtering out inconsistent scoring results; The verified injury score results are merged, and the highest score of multiple injury scores in the same part is taken as the score of the part, and the highest score of multiple part scores in the same area is taken as the score of the area; determine whether the score of each area is 6 points. If so, the total ISS score is determined to be 75 points. If not, select the three highest area scores, and take the sum of the squares of these three area scores as the total ISS score; output the total ISS score and the specific score results of each trauma area.

2. The method according to claim 1, characterized in that The artificial intelligence model is guided by the prompt word engineering to identify the diagnostic information related to the trauma in the structured medical record data, and the detailed description related to the diagnostic information is extracted from the structured medical record data, and the JSON format data containing the diagnostic information, injury identification and injury description is generated, including: A prompt word template is constructed based on a four-layer structure of role-goal-process-format, wherein the role in the prompt word template is set as an emergency doctor, the goal in the prompt word template is set as identifying trauma-related diagnoses and extracting detailed description information, the process in the prompt word template is set as diagnosis extraction, relevance judgment and description extraction, and the format in the prompt word template is set as JSON format; A multidimensional analysis method based on clinical characteristics is used to judge the relevance of trauma, which includes analyzing the time characteristics of diagnosis, judging whether it is caused by external force according to the injury mechanism, judging whether it meets the characteristics of traumatic injury in combination with the anatomical location and clinical manifestations, and establishing a trauma feature word library for feature word matching; Detailed descriptions related to diagnosis were extracted based on the contextual association analysis method, by locating the occurrence of injuries in the current medical history, extracting injury signs from physical examination records, and extracting imaging manifestations and laboratory test results from auxiliary examination results; a hierarchical extraction strategy was used to extract directly related descriptive information, indirectly related clinical manifestations, and objective examination results in turn; The extracted information is converted according to a three-layer JSON template, which includes: a diagnosis information layer for storing standardized diagnosis names, an injury identification layer for indicating the relevance of the diagnosis to the current trauma, and an injury description layer for organizing detailed information according to the dimensions of current medical history, physical examination, and auxiliary examinations; a medical terminology standardization mapping table is established to convert clinical expressions into standard medical terms, and an information integrity check mechanism is used to ensure that each diagnosis has the necessary descriptive information.

3. The method according to claim 1, characterized in that The scoring rules related to a single injury are retrieved from the adapted ISS scoring rule knowledge base using vector retrieval technology, including: The ISS scoring rules are structured and adapted, and the ISS scoring rules are classified according to the head and neck, face, chest, abdomen, limbs and body surface areas. A scoring hierarchy is established in each area according to the degree of injury. Each scoring rule is described in a standardized manner, and the standardized description includes key anatomical structures, injury types, and injury degrees, and is supplemented with common clinical variant descriptions and clinical equivalent descriptions to form a scoring rule knowledge base; The scoring rules in the scoring rule knowledge base are vectorized by using a pre-trained language model in the medical field, and each scoring rule is decomposed into three semantic segments: anatomical location, injury type, and severity. The three segments are encoded and fused to generate a rule vector representation. The rules in the same anatomical region with an injury degree less than a preset injury threshold are used as positive sample pairs, and the rules in different regions or with an injury degree greater than the preset injury threshold are used as negative sample pairs through a comparative learning mechanism, so as to minimize the vector distance between the positive sample pairs and maximize the vector distance between the negative sample pairs. A multi-level vector retrieval mechanism was constructed. Coarse-grained filtering based on anatomical regions was performed to narrow the retrieval scope. An approximate nearest neighbor retrieval index was constructed using a vector indexing algorithm. Cosine similarity was used for precise matching measurement within the candidate set. The similarity threshold was adaptively adjusted according to the rule complexity of different anatomical regions. The matching results are sorted based on vector similarity, and specific rules are given priority considering rule specificity. A rule credibility scoring mechanism is established in combination with clinical verification data to screen the most suitable scoring rules. When it is impossible to directly match a completely compliant rule, scoring recommendations are derived by analyzing the scoring patterns of similar rules and combining the principles for judging the severity of the injury.

4. The method according to claim 1, characterized in that: Construct a multi-level vector retrieval mechanism, perform coarse-grained filtering based on anatomical regions to narrow the retrieval scope, use vector indexing algorithm to construct an approximate nearest neighbor retrieval index, use cosine similarity for precise matching measurement in the candidate set, and adaptively adjust the similarity threshold according to the rule complexity of different anatomical regions, including: An anatomical region index table is established to divide the human body into the head and neck, face, chest, abdomen, limbs and body surface regions, anatomical structure sub-regions are set in the anatomical region index table, scoring rules are mapped to corresponding anatomical regions through an inverted index, and the relevant regions are located in the anatomical region index table based on the anatomical position information in the input injury to obtain a first candidate rule set; A hierarchical graph structure is used to construct a vector index for the first candidate rule set. The total number of rules in the first candidate rule set is recorded as N, and the number of hierarchical layers is set to log(N). An approximate nearest neighbor graph is constructed in each layer as a sparse subset of the lower layer graph. The maximum out-degree of each layer node is set to 16 to control the density of the graph. After randomly selecting an entry point at the top layer for each rule vector in the first candidate rule set, a greedy search is performed on each layer to find the most similar node to establish a connection to form an index structure. The second candidate rule set is retrieved based on the index structure. The similarity between the rule vectors in the second candidate rule set and the input injury query vector is calculated, and the cosine similarity is used as the basic measurement indicator to calculate the similarity between the rule vector and the query vector. The position weight factor is set according to the importance of the anatomical structure, and the weight of organ injury is set higher than the weight of soft tissue injury. The weighted similarity score is calculated as a quantitative indicator of rule matching to obtain the rule set to be matched, and the similarity threshold is adaptively adjusted according to the complexity of the rules in different anatomical regions.

5. The method according to claim 1, characterized in that The artificial intelligence model identifies the region and part to which each injury belongs based on the diagnostic information, injury identification and injury description in the JSON format data, combined with the retrieved scoring rules, and determines the corresponding injury scoring results including: Pre-training a medical text corpus to obtain a pre-trained model, wherein the medical text corpus includes diagnostic reports, surgical records, and medical record writing specifications, and using a masked language model task to focus on training the semantic representation of anatomical site words and injury description words, setting the pre-trained model to a dual-stream attention structure, wherein the first attention stream is used to analyze text context semantic relations, and the second attention stream is used to identify medical entity associations, and the pre-trained model outputs a medical concept feature vector; Based on the medical concept feature vector, an anatomical knowledge graph layer is constructed, wherein the anatomical knowledge graph layer includes an anatomical knowledge base and a relational reasoning network, wherein the anatomical knowledge base uses a tree structure to store standardized human anatomical terms, and the relational reasoning network uses a graph attention network to implement message transmission, and the medical concept feature vector is mapped and aligned with the anatomical terminology to generate an anatomical feature mapping result; A hierarchical anatomical structure classifier is constructed according to the anatomical feature mapping results, the anatomical feature mapping results are divided into anatomical regions using a sequence annotation model based on a bidirectional long short-term memory network, the anatomical entity boundaries are identified using a biological information annotation scheme to obtain anatomical region annotation results, a fine-grained classification model is constructed for the anatomical region annotation results, and the features of the anatomical region annotation results are transferred to the fine-grained classification model through a residual connection mechanism to obtain an anatomical structure classification result; A semantic similarity calculation model is constructed based on the anatomical structure classification result. The semantic similarity calculation model uses a Transformer structure to vectorize the injury description and the standard description of the rule library. A correlation model between feature words and scoring results is established by analyzing historical scoring data. The gradient descent method is used to optimize the feature word weight coefficient to obtain the feature word weight vector. The feature word weight vector and the candidate rules are constructed into a directed graph structure. The rule combination pattern is learned through graph convolution operation to obtain the rule matching result. The attention pooling mechanism is used to aggregate the multiple trauma scores of the rule matching results, and the final injury score result is output.

6. The method according to claim 5, characterized in that By analyzing the historical scoring data, a correlation model between feature words and scoring results is established. The feature word weight coefficient is optimized by the gradient descent method to obtain a feature word weight vector. The feature word weight vector and the candidate rules are constructed into a directed graph structure. The rule combination pattern is learned through graph convolution operation to obtain the rule matching results including: Construct a feature word score mapping matrix, the feature word score mapping matrix includes feature word occurrence frequency vectors and corresponding score values, adopt a stratified sampling method to obtain sample sets with different score intervals, calculate the discrimination of feature words in high-score samples and low-score samples based on the sample sets to obtain initial correlation weights, and generate sample subsets through a bootstrap resampling method to calculate the average value of correlation coefficients; A feature word co-occurrence network is constructed based on the average value of the correlation coefficient, the network centrality of the feature word co-occurrence network is calculated, the network centrality is combined with the average value of the correlation coefficient to obtain a feature word weight vector, the feature word weight vector is iteratively optimized using a stochastic gradient descent method with a momentum term, the mean square error between the predicted score value and the true score value is calculated as a loss function in each iteration, and an L2 regularization term is introduced to avoid overfitting; A rule matching directed graph is constructed using the feature word weight vector, wherein the vertex set of the rule matching directed graph includes candidate rule nodes and feature word nodes, the edge set of the rule matching directed graph represents the association relationship between the rule and the feature word, and a multi-head attention mechanism is used to calculate the edge weight between the rule node and the feature word node, wherein the edge weight is jointly determined by the query vector of the rule node and the key vector and value vector of the feature word node; A graph convolutional network is applied to the rule matching directed graph, and layer-by-layer abstraction of rule node representation is achieved by stacking multiple layers of graph convolution. Each layer of graph convolution aggregates and updates the neighbor information of the node through a learnable weight matrix to obtain a high-level representation of the rule node. After the last layer of graph convolution, an attention readout mechanism is introduced to calculate the contribution of different rule nodes, and a weighted voting method is used to combine rule nodes with higher contributions to obtain the rule matching result.

7. A trauma severity assessment system based on artificial intelligence and big data analysis, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to extract structured medical record data of trauma patients from the trauma treatment information system, wherein the structured medical record data includes current medical history, physical examination, auxiliary examination and diagnosis information; guide the artificial intelligence model to identify the diagnosis information related to trauma in the structured medical record data through prompt word engineering, and extract detailed descriptions related to the diagnosis information from the structured medical record data, and generate JSON format data including diagnosis information, injury identification and injury description; The second unit is used to score each injury separately based on the split scoring strategy, specifically including: using vector retrieval technology to retrieve scoring rules related to a single injury from the adapted ISS scoring rule knowledge base, wherein the ISS scoring rule knowledge base uses markdown language to adapt the original ISS scoring rules; the artificial intelligence model identifies the area and part to which each injury belongs based on the diagnostic information, injury identification and injury description in the JSON format data, combined with the retrieved scoring rules, and determines the corresponding injury scoring results; the injury scoring results are verified, and inconsistent scoring results are filtered out by comparing whether the injury name in the injury scoring results is consistent with the input diagnostic information; The third unit is used to merge the verified injury score results, take the highest score of multiple injury scores in the same part as the score of the part, and take the highest score of multiple part scores in the same area as the score of the area; determine whether the score of each area is 6 points. If so, the total ISS score is determined to be 75 points. If not, the three highest area scores are selected, and the sum of the squares of the three area scores is used as the total ISS score; output the total ISS score and the specific score results of each trauma area.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images